Related Experiment Video
Updated: May 29, 2026
06:49
Evaluation of a Smartphone-based Human Activity Recognition System in a Daily Living Environment
Published on: December 11, 2015
Evaluation of artificial neural network algorithms for predicting METs and activity type from accelerometer data:
Patty S Freedson1, Kate Lyden, Sarah Kozey-Keadle
1Dept. of Kinesiology, University of Massachusetts, Amherst, Massachusetts, USA. psf@kin.umass.edu
Journal of Applied Physiology (Bethesda, Md. : 1985)
|September 3, 2011
Summary
Artificial neural networks (nnets) accurately estimate physical activity intensity and type from accelerometer data. This machine learning approach shows promise for broader applications in health and exercise research.
Area of Science:
- Biomedical Engineering
- Kinesiology
- Machine Learning
Background:
- Previous research demonstrated the potential of artificial neural networks (nnets) for estimating metabolic equivalents (METs) and activity type using accelerometer data.
- The accuracy and generalizability of these initial models were limited by the training dataset size and diversity.
Purpose of the Study:
- To develop and validate enhanced artificial neural networks (nnets) using a larger, more diverse dataset for improved estimation of physical activity.
- To assess the robustness and flexibility of machine learning techniques when applied to independent validation samples.
Main Methods:
- Trained nnets using data from 277 participants completing 11 activities (University of Massachusetts).
- Validated models on an independent sample of 65 participants (University of Tennessee) performing three activity routines.
- Utilized open-circuit indirect calorimetry for criterion MET measurement and direct observation for activity type classification.
- Input variables included accelerometer count distribution features and lag-1 autocorrelation.
Main Results:
- The nnet demonstrated a bias of +0.32 METs and a root mean square error of 1.90 METs when applied to the independent sample.
- Seventy-seven percent of activities were correctly classified into sedentary/light, moderate, or vigorous intensity categories.
- Activity type classification accuracy was high for household (98.1%) and locomotion (89.5%) activities, but lower for sports (23.7%).
Conclusions:
- Machine learning models, specifically nnets, show reasonable performance when applied to independent datasets for physical activity assessment.
- The findings support the potential of using accelerometer data and nnets for objective physical activity monitoring.
- Establishing an open-access activity dictionary with diverse accelerometer data is proposed to further enhance prediction accuracy for METs, activity intensity, and type.
